Skip to content
Bud's Wiki

Few-Step Generation (from scratch)

TL;DR

Section titled “TL;DR”

Few-step generation covers one- and few-step (1–4 NFE) generative models trained from scratch — not via teacher-student distillation. The wiki’s filed instances so far split along the modality axis: SoFlow (SoFlow: Solution Flow Models for One-Step Generative Modeling) targets image generation via a bi-time solution-function parameterization with a JVP-free Taylor-consistency loss; DriftWorld (DriftWorld: Fast World Modeling through Drifting) applies the Kaiming-lab drifting paradigm to action-conditioned video for robot world modeling; A2A (A2A: Action-to-Action Flow Matching) shortens the flow-matching trajectory itself for robot action heads by anchoring the “noise” endpoint to the previous proprioceptive action. Common substrate: compress the multi-step denoising trajectory into a single (or handful of) forward passes by design, not by post-hoc distillation of a longer teacher.

Key claims

Section titled “Key claims”
  • From-scratch few-step is a distinct research axis from distillation. Distillation compresses a pretrained multi-step teacher into a few-step student and is tracked separately on Diffusion Distillation with 95+ filed papers (DMD, TDM, ArcFlow, RMD, DiagDistill, …). From-scratch few-step never had a step count to reduce — the trajectory is designed to be short — and the training objective is what does the lifting (SoFlow: Solution Flow Models for One-Step Generative Modeling §3).
  • JVP-based few-step training is a scaling bottleneck. MeanFlow’s self-consistency requires Jacobian-vector products that are poorly optimized in PyTorch; SoFlow’s Taylor-around-nearby-times consistency cancels the JVP algebraically, giving strictly better 1-NFE FID at every DiT scale (B/2 4.85 vs 6.17, XL/2 2.96 vs 3.43 on ImageNet 256×256) (SoFlow: Solution Flow Models for One-Step Generative Modeling Table).
  • Training-time Classifier-Free Guidance is rare for from-scratch one-step models. SoFlow’s Flow Matching branch regresses against a guided marginal velocity target during training, letting a one-step model benefit from CFG without the two-pass inference overhead (SoFlow: Solution Flow Models for One-Step Generative Modeling §3).
  • The drifting paradigm extends from images to action-conditioned video. DriftWorld reports 30+ FPS on a single H100 (17× vs diffusion-based WFM baselines) across Bridge-V2 / RT-1 / Language Table / Robomimic / Push-T, and 1–2 GPUs is enough to train the backbone — concrete evidence that teacher-free one-step generation scales to sequence-output settings beyond ImageNet (DriftWorld: Fast World Modeling through Drifting §Visualization).
  • For domains with strong temporal continuity, from-scratch few-step can be achieved by shortening the trajectory itself. A2A observes that adjacent robot action chunks are physically similar, so anchoring the flow’s “noise” endpoint to the previous proprioceptive action makes the trajectory short enough that a single Euler step suffices — 0.56 ms latency at ≥90% task success across five Robosuite tasks (A2A: Action-to-Action Flow Matching §Ingredient 2).
  • Robustness gains, not just latency gains. A2A reports the largest surviving success margins under Level-3 visual perturbations (38% at 6 NFE / 32% at 1 NFE vs FM-UNet 16% at 10 NFE / DDPM-UNet 10% at 100 NFE) and 80% real-world OOD on a glowing-cube variant of Pick Cube where FM-UNet and DDPM-UNet collapse to 0% — evidence that the from-scratch few-step design is not a Pareto trade against generalization (A2A: Action-to-Action Flow Matching Visual Generalization Results table).

Recent contributions

Section titled “Recent contributions”
  • [2026-11-05] ODE Policy: Robot Learning via Action Dynamics, Not Actions: ODE Policy (Jia et al.; NTU MARS Lab) adds a fifth from-scratch few-step recipe distinct from SoFlow’s bi-time solution function, DriftWorld’s drifting, A2A’s proprioceptive-endpoint anchoring, and Explorative Modeling’s K-candidate exploration: parameterize the vector field dā/dτ = v_θ(ā, τ, o) directly and let a numerical ODE solver integrate each executed action from the previous one, so there is no multi-step denoising trajectory to compress. Reports 10.8× faster action generation than flow matching under matched one-epoch IL budgets with 5.9× the success rate, and ~4× speedup as a VLA/WAM head (π0.5 96.0% vs 75.5%, 4.5B WAM 82.6% vs 24.6%, both at one epoch). First filed instance on this page where the few-step-generation recipe is applied to robot action heads at scale rather than at the chunk-anchor level A2A: Action-to-Action Flow Matching A2A targets — and the first where the structural-continuity-by-construction property is explicitly reported as yielding smoother-than-human trajectories without an auxiliary smoothness loss. A direct test of the page’s second open question (is “shorten the flow between adjacent latents” a general recipe): ODE Policy’s answer is that once the parameterization is the vector field itself, the proprioceptive-anchor trick becomes a structural property rather than a sampler-side choice, and the 3.9× mode-balance gain over flow matching answers A2A’s robustness-vs-multimodality open question in the robot-action regime.
  • [2026-10-16] Are We Overthinking One-Step Generation?: Priming Models (Park, J. Kim, Du, Chen, Han) propose the simplest from-scratch one-step recipe filed so far: reconstruction pretraining at a single fixed noise level (no time embedding, no teacher, no CFG, no JVP), then feature-space distributional post-training. Reaches ImageNet 256² one-step FID 1.19 at 6.2× fewer FLOPs than Drifting-L and 12.2× fewer than Drifting-B — a direct FLOPs-and-FID win against the Drifting lineage underpinning DriftWorld. Scales to T2I (GenEval 0.818 one-step vs 0.548 for one-step MeanFlow) and robot world modeling (RT-1 / Bridge-V2). Fourth distinct from-scratch recipe on the page (SoFlow’s solution function, DriftWorld’s drifting, Explorative Modeling’s K-candidate exploration, now Priming’s reconstruction-first). Shares the “relies on pretrained feature encoders” limitation with Drifting, keeping that open question alive.
  • [2026-10-16] Residual Modeling Closes the Regression and Generative Policy Gap in Robot Learning: HT-Policy (Zhou et al., Tsinghua + MIT) pushes the “zero-step generation via regression” extreme of the step-count axis in the robot-policy regime: a single-pass heteroscedastic Student-t regressor on existing flow-matching VLA backbones (GR00T N1.7, π0.5, Cosmos3-Nano) matches flow-matching baselines at task success while running 1.82–17.83× faster. The analysis is a diagnostic complement to few-step distillation (A2A, DriftWorld, SoFlow): rather than compressing a learned velocity field into fewer Euler steps, HT-Policy argues the velocity field is a workaround for MSE’s heavy-tail-gradient problem — fix the loss and the regressor is competitive at zero denoising steps. Direct contrast with A2A: Action-to-Action Flow Matching A2A’s starting-distribution lever and the broader flow-matching few-step line in the robotics regime: HT keeps the flow-matching backbone weights but swaps the inference-time generation procedure for a feed-forward pass.
  • [2026-10-13] Velocity Scaling in Flow Matching: Velocity Scaling in Flow Matching (Saied & Fleuret) is an inference-only one-multiplication intervention that recovers much of the quality gap at low NFE without any retraining or distillation — on SiT-XL/2 ImageNet-256 at NFE 25, FID drops 28.0 → 12.2 (lag correction) → 8.7 (extra sharpening). Orthogonal to the from-scratch few-step recipes on this page (SoFlow, DriftWorld, A2A, Explorative Modeling): applies to an already-trained standard flow-matching model and is a non-distillation alternative at the inference side.
  • [2026-07-31] Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation: Explorative Modeling adds a fourth from-scratch few-step recipe distinct from SoFlow’s bi-time solution function, DriftWorld’s action-conditioned drifting, and A2A’s proprioceptive-endpoint anchoring: end-to-end reconstructive XMs use a single-forward-pass generator (no step decomposition at all) trained with K-candidate exploration, and match diffusion on control tasks with 16–256× fewer inference steps. Notable because it’s the first filed instance where the training-side load (K parallel candidates + argmin) is what enables the inference-side collapse to one step, rather than a solution-function parameterization (SoFlow) or a shortened trajectory (A2A).
  • [2026-07-24] A2A: Action-to-Action Flow Matching: A2A (RSS 2026) is the first filed instance of the “shorten the flow trajectory itself” recipe for from-scratch few-step action generation — initialize the flow-matching ODE at a lifted embedding of the previous proprioceptive action instead of Gaussian noise, so adjacent-chunk continuity carries the load and one Euler step suffices. Also reports Level-3 visual generalization margins and 80% real-world OOD on Pick Cube variants, arguing the design compounds robustness with latency rather than trading against it.
  • [2026-07-21] DriftWorld: Fast World Modeling through Drifting: DriftWorld extends the Kaiming-lab drifting paradigm to action-conditioned video for robotics WFMs. 30+ FPS on a single H100 across 5 benchmarks (17× vs diffusion-based baselines), 1–2 GPUs to train, and the demo runs on an NVIDIA L4. First filed application of drifting to a sequence-output WFM.
  • [2026-05-22] SoFlow: Solution Flow Models for One-Step Generative Modeling: SoFlow trains a bi-time solution-function f_θ(x_t, t, s) directly with Flow Matching + a JVP-free Taylor consistency loss, beating MeanFlow at every DiT scale on ImageNet 256×256 at 1 NFE (XL/2 2.96 vs 3.43). Also supports training-time CFG, an efficiency lever rare in the from-scratch one-step regime.

Open questions

Section titled “Open questions”
  • Do these three recipes compose, or are they mutually exclusive design points? SoFlow’s solution-function parameterization, DriftWorld’s drifting objective, and A2A’s proprioceptive-endpoint anchoring each address a different failure mode of naive one-step generation. Whether they can be stacked (e.g. drifting + proprioceptive anchoring for a robot WFM) is untested.
  • Is the “adjacent latents are close, so the flow between them is short” principle a general recipe? A2A exploits proprioceptive continuity for actions; Causal Forcing / DiagDistill exploit chunk continuity for AR video (Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation, Streaming Autoregressive Video Generation via Diagonal Distillation — both distillation-side). Whether the same idea gives a from-scratch few-step AR video model, or generalizes beyond temporally continuous domains, is open.
  • Does from-scratch few-step scale to production-quality video? All three filed instances stop short of the ~14B-parameter Wan-class regime where distillation-based recipes (TMD, RMD, DiagDistill) hold the current frontier. Whether SoFlow’s solution-function or DriftWorld’s drifting can match distillation at that scale — or whether the from-scratch recipes stay on the small-model / specialized-task end of the Pareto — has not been settled.
  • What is the right marginal-KL anchor for from-scratch few-step models? TDM-R1 uses a pretrained TDM teacher as its anchor for RL post-training; from-scratch few-step generators have no such anchor. Either the recipe needs a teacher-free regularizer (the generator’s own EMA?) or RL post-training doesn’t compose cleanly with this recipe family.
  • Is CFG-during-training generalizable beyond SoFlow? DriftWorld and A2A do not report CFG results at all. Whether the SoFlow trick of regressing against a guided marginal velocity target transfers to the drifting objective (where there is no explicit velocity field) or to the anchored-endpoint setting (where the endpoint distribution is degenerate) is open.

Papers

Section titled “Papers”

2026-07

Section titled “2026-07”

2026-01

Section titled “2026-01”

2025-12

Section titled “2025-12”